In partnership with

Good morning. It’s Monday, October 27th.

On this day in tech history: In 1994, Intel and Nestor Corp introduced the Ni1000, the first all-digital neural-network chip designed for image analysis, capable of processing 40,000 pattern recognitions per second using over 3 million transistors.

In today’s email:

• OpenAI Developing Music Generator?

• Gemini’s Visual Layout Feature

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

OpenAI developing AI music generator as Altman questions what counts as ‘real work’

OpenAI is reportedly building a generative music model that can compose or enhance tracks from text or audio prompts, marking its most ambitious creative push since 2020’s Jukebox. The system can generate full compositions or add specific instrumental layers, like guitar or percussion, based on user direction.

To refine musical structure and phrasing, OpenAI has enlisted Juilliard students to annotate sheet music used in training. The model draws on prior breakthroughs in text-to-speech and sound synthesis and could be integrated into ChatGPT or Sora’s media pipeline.

Speaking with Rowan Cheung at DevDay, Altman said work soon to be overtaken by AI isn’t “real work.” He argued that technological progress continually redefines what counts as valuable effort: “A farmer 50 years ago would look at what you and I do and say, ‘that’s not real work.’ Altman’s comment, both fatalistic and forward-looking, suggested that AI might erase millions of jobs yet also create new, less tangible forms of productivity.

The company’s Startup Fund also recently backed Valthos, a biodefense venture using machine learning to detect pathogens and update vaccines in real time. Founded by former Palantir and Oxford researchers, Valthos claims its models can compress response times from months to hours, an advance OpenAI calls vital for “national resilience.” Read more.

First look at Google’s Gemini Visual Layout feature where AI responses turn into live dashboards

Google’s upcoming Gemini Visual Layout could be the most visible leap yet. Instead of static text, you’ll get a modular feed of cards packed with tables, sliders, and interactive visuals. It feels more like a live dashboard than a chat window. It’s a shot across the bow of OpenAI’s Pulse and a signal that Google wants Gemini to handle serious workflow logic before its long-awaited “Agent Mode” automates execution.

Over at DeepMind, the new Dreamer 4 agent is teaching itself to think in worlds that don’t exist. It learns entirely inside a simulated “world model,” predicting what would happen instead of trying things for real. It trained on Minecraft gameplay videos and became the first AI to mine diamonds without ever playing the game. A transformer core predicts actions and rewards, while a “shortcut forcing” technique makes generation 25 times faster than conventional video-based learning. DeepMind says that same tech could help train robots safely in simulation before they ever touch hardware.

Google’s physical world ambitions are just as bold. The company is backing a 400-megawatt carbon-capture power plant in Illinois to power its data centers while trapping up to 90 percent of its emissions. It’s a big bet on a technology that’s still hit-or-miss, and one that critics say can’t offset the methane leaks that come with burning natural gas.

And in a sign of where this is heading, Google’s own AI team is teasing a future where coding feels more like collaboration than syntax. Developers may soon “vibe code,” describing video games and tools in natural language while AI builds them in real time. Read more.

Researchers pinpoint three keys to making AI agents far smarter 

Researchers from the National University of Singapore, Princeton, and Illinois Urbana-Champaign have identified three key levers that make AI agents significantly smarter: data quality, algorithm design, and reasoning style. Their findings show that a well-trained 4-billion-parameter model can match or even outperform competitors with up to 32 billion parameters, highlighting that training strategy can outweigh sheer scale.

Data quality proved critical. Models trained on authentic, continuous learning trajectories captured full reasoning workflows, including planning, guided execution, error correction, and tool integration, while those trained on synthetic data underperformed by a wide margin. Algorithm design also made a difference: token-level scoring combined with broader exploration and optimized rewards (GRPO-TCR) improved both stability and accuracy compared with conventional reinforcement learning approaches.

Finally, reasoning strategy mattered: deliberative models that plan actions carefully outperformed reactive models that rely on fast, repeated tool calls. The resulting DemyAgent-4B achieved competitive results across math, science, and coding benchmarks. Both the model and training datasets are publicly available. Read more.

• Microsoft Outlook is getting an AI overhaul under new leaders

• Meta AI preps major feature push as TBD Lab takes the reins

• A 300,000-scenario experiment shows how Anthropic, OpenAI, and Google’s AIs disagree on the same moral questions

• Your LinkedIn activity could soon train its AI, unless you opt out by November 3

• An Ohio lawmaker just introduced a bill banning AI marriages and declaring chatbots officially “non-sentient”

• Dropbox rolls out Dash, an AI assistant that actually understands your files

• New study finds leading AI models spiral into compulsive betting — and can’t stop until they lose it all

• Bonsai Robotics’ new Amiga vehicles use vision-based AI to make farms fully autonomous

• Carnegie Mellon built a drone that 3D-prints structures from text prompts

• AI models can learn a writer’s voice from just two books. Even experts can’t tell the difference

• Study finds AI chatbots cite obscure websites that never make Google’s top results

• Misleading claims that most web content is now AI-generated are spreading on social media

• AI policy without proof is just politics, says Berkeley professor calling for independent oversight

• Inside Ring-1T: Ant engineers solve reinforcement learning bottlenecks at trillion scale

• Trump and Xi to finalize TikTok pact, marking a truce in the global AI race

• Elon Musk says he won’t build Tesla’s ‘robot army’ unless he keeps control of it

• The impact of AI on the future job market may include some 'transitional friction'

• A famed economist warns Gen Z will have to “work a lot harder” to survive the AI age, and still might end up poorer

• Sakana AI's CTO says he's 'absolutely sick' of transformers, the tech that powers every major AI model

• The glaring security risks with AI browser agents

• Why quantum computing needs a NASA-like movement

• AI models may be developing their own ‘survival drive’, researchers say

• Mistral launches its own AI Studio for quick development with its European open source, proprietary models

📄 Every Question Has Its Own Value: Reinforcement Learning with Explicit Human Values

📄 DyPE: Dynamic Position Extrapolation for Ultra High Resolution Diffusion ​​

📄 Open-o3 Video: Grounded Video Reasoning with Explicit Spatio-Temporal Evidence  

📄 AdaSPEC: Selective Knowledge Distillation for Efficient Speculative Decoders

📄 HoloCine: Holistic Generation of Cinematic Multi-Shot Long Video Narratives

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!